"I always figured that voltage was the primary determinant of how two otherwise identical processors performed from a power perspective, but it turns out I was wrong."
No you were right! Either the chips will perform very similarly when clocked and powered the same, or those chips are not identical :)
"Either the chips will perform very similarly when clocked and powered the same, or those chips are not identical :)"
From a pragmatic point of view you're absolutely right. However, there are at least two kinds of "identical" here: logical and physical. In my post above I talked a bit about how the chips could be logically identical but still differ a lot physically, despite using the same stepping and/or whatever.
If the QX6950 won't work then this new hotter beast sure wont. Why is no one investigating why the Yorkfields that we where assured to work on 680i motherboards will NOT work now???
They will work on 680i but probably not too well overclock... But perhaps, you should ask nvidia why they prefer sell news chipsets 780i than support already sold one...
[quote]"I always figured that voltage was the primary determinant of how two otherwise identical processors performed from a power perspective, but it turns out I was wrong.[/quote]
If the chips were indeed identical, you would still be right (for a single chip voltage is the primary determinant). However, if you consider the small structures we are dealing with, gate oxide thicknesses of 1.2 nm / a few mono layers (!), and the large influence which they can have on device performance (gate leakage current depending exponentially on gate oxide thickness), it can be seen that:
- it's inevitable to have some variation within each chip, between the chips on each wafer and between the batches (won't go into the reasons now)
- the smaller the structures get and the more transistors we pack into each chip, the more pronounced the effect of such variations becomes
The hafnium-based high-k oxide in Intels 45 nm process enables them to use thicker oxide and achieve the same electrical performance, so gate leakage should not be the problem here. But there are many more transistor parameters to consider and other leakage paths (substrate to gate, subthreshold current).
Having said that I can imagine that Intel:
- may have used a thinner gate oxide intentionally to increase frequency headroom, but got unexpected results from the rather new material system "Hf-oxide + metal gate" (higher leakage due to a higher defect density in the oxide, tunneling via Poole-Frenkel emission)
- may have gotten a "normal" variation in the process and, as you suggested, your X9650 may have been somewhat lucky
- may have gotten an unintentional variation in some process parameters
- may come up with some, for ousiders, totally unexpected explanation
So, depending on what Intel tells you, it may be a good idea to keep an eye on the variation of power consumption between chips from different batches of Intels 45 nm process.
Has Intel said anything about how they'll be handling mounting for the new socket Nehalem will use? If they were going to change the mounting system, that would be their chance.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
8 Comments
Back to Article
Luminair - Monday, November 26, 2007 - link
"I always figured that voltage was the primary determinant of how two otherwise identical processors performed from a power perspective, but it turns out I was wrong."No you were right! Either the chips will perform very similarly when clocked and powered the same, or those chips are not identical :)
MrSpadge - Monday, November 26, 2007 - link
"Either the chips will perform very similarly when clocked and powered the same, or those chips are not identical :)"From a pragmatic point of view you're absolutely right. However, there are at least two kinds of "identical" here: logical and physical. In my post above I talked a bit about how the chips could be logically identical but still differ a lot physically, despite using the same stepping and/or whatever.
Stephan
aguilpa1 - Wednesday, November 21, 2007 - link
If the QX6950 won't work then this new hotter beast sure wont. Why is no one investigating why the Yorkfields that we where assured to work on 680i motherboards will NOT work now???nemrod - Wednesday, November 21, 2007 - link
They will work on 680i but probably not too well overclock... But perhaps, you should ask nvidia why they prefer sell news chipsets 780i than support already sold one...MrSpadge - Wednesday, November 21, 2007 - link
[quote]"I always figured that voltage was the primary determinant of how two otherwise identical processors performed from a power perspective, but it turns out I was wrong.[/quote]If the chips were indeed identical, you would still be right (for a single chip voltage is the primary determinant). However, if you consider the small structures we are dealing with, gate oxide thicknesses of 1.2 nm / a few mono layers (!), and the large influence which they can have on device performance (gate leakage current depending exponentially on gate oxide thickness), it can be seen that:
- it's inevitable to have some variation within each chip, between the chips on each wafer and between the batches (won't go into the reasons now)
- the smaller the structures get and the more transistors we pack into each chip, the more pronounced the effect of such variations becomes
The hafnium-based high-k oxide in Intels 45 nm process enables them to use thicker oxide and achieve the same electrical performance, so gate leakage should not be the problem here. But there are many more transistor parameters to consider and other leakage paths (substrate to gate, subthreshold current).
Having said that I can imagine that Intel:
- may have used a thinner gate oxide intentionally to increase frequency headroom, but got unexpected results from the rather new material system "Hf-oxide + metal gate" (higher leakage due to a higher defect density in the oxide, tunneling via Poole-Frenkel emission)
- may have gotten a "normal" variation in the process and, as you suggested, your X9650 may have been somewhat lucky
- may have gotten an unintentional variation in some process parameters
- may come up with some, for ousiders, totally unexpected explanation
So, depending on what Intel tells you, it may be a good idea to keep an eye on the variation of power consumption between chips from different batches of Intels 45 nm process.
Regards, Stephan
IntelUser2000 - Thursday, November 22, 2007 - link
Problem is that other review sites aren't showing such a disparity between the QX9650 and QX9770 and some like Neoseeker does: http://www.legitreviews.com/article/596/11/">http://www.legitreviews.com/article/596/11/MrSpadge - Thursday, November 22, 2007 - link
"Funny" about Neoseeker is that they fail to recognize this CPU chewing ~50W more for a ~7% higher clock speed as unusual.Stephan
ViRGE - Wednesday, November 21, 2007 - link
Has Intel said anything about how they'll be handling mounting for the new socket Nehalem will use? If they were going to change the mounting system, that would be their chance.